• AI
  • Story

NVIDIA's free AI API is fine for testing, not for your private data

My take: Free to test is a fair deal. But what I send may train NVIDIA's models, so nothing private goes there: that's what a model at home is for.

What to know

  • On November 16, 2026, NVIDIA's catalog listed 40 models with a free endpoint, among them DeepSeek V4.1 Flash and GLM 5.3 Flash. A model's page gives you an API key (the secret a program sends to prove it may call the service). Its sample code uses OpenAI's client with NVIDIA's address, so a tool that accepts an OpenAI-compatible address and key can use it.
  • The list moves. The same day, GLM 5.3 and Kimi K3, both shown in the reel, were marked deprecated for the free endpoint, and the page of Kimi K3.1 Pro said its API was not open yet. NVIDIA publishes no fixed quota: its FAQ says the rate varies with the model and the number of users, and shows yours in your account.
  • The trial terms (version of September 19, 2025) allow internal testing and evaluation, not production. You agree not to send confidential or personal data, and NVIDIA may use what you send and what the model answers, without identifying you, to improve its products, AI models included.

Spotted on @jacky_west's reel, in French.

What changes

For you if you want to try large open models in a coding agent without paying per request or owning a big GPU: one free key lets you compare several on your own project. The catch is the data. What you paste leaves for NVIDIA's servers, under terms that let it learn from it, and a model can lose its free endpoint from one month to the next. Use it for public code and test prompts; for anything private, run an open model on your own machine. Create a key, try two models on a side project, and check your quota before you build on it.

Source: NVIDIA, “NVIDIA API Trial Terms of Service”, September 19, 2025

Drafted with Claude, Anthropic's AI, from my take. Reviewed and owned by me.

Up next